Commit Graph
950 Commits
Author SHA1 Message Date
Willie c0e66720ce fix(source-links): preserve trashed links and resolve audit and hash roots
Trash moves and restores the link entry without copying or removing its target, including when rename fails. Relative link text is stored as an absolute target so relocation cannot change its meaning; list and TUI expose the link target without inventorying its contents.

Pass the operation-scoped Follow policy into audit and hash operations, resolving their physical root while keeping logical result and metadata keys. Calls without a policy keep their existing behavior and ScanSkill retains its original function signature.

Update six stale trash raw-write allowances for the injected rename test seam and classify moveLink creation/removal: it moves a checked source entry to trash or restores an entry into source, never writes through its target. Raw sourcewalk calls and allowances are unchanged.
2026-10-05 14:45:41 +08:00
Willie a502e9316b feat(sourcefs): write through followed first-level source links
With follow_source_links on, discovery reports <source>/_dev-skills/foo,
but every write to it was refused: sourcefs rejects a link in the path, and
its source-root os.Root refuses descendants outside the source. Dashboard
edits, nested uninstall, regular skill replacement, and tracked installs
under such a link all failed, although writing through a link the user
placed is the agreed behavior (#410).

sourcefs.Open, Create, MkdirAllIn, and CheckMoveOut take the operation's
optional sourcewalk.Follow. A clean name <link>/<rest> whose first
component the policy follows is handled by a second os.Root opened at the
link's canonical target, with the same per-component link check on <rest>,
so a nested link or ".." still cannot escape the target. Names with ".."
keep the source root, which refuses them through the link. Without a
policy nothing changes.

The link entry itself stays a link: RemoveAll, Rename, and writes on it are
refused, install --track --force still refuses to replace it, and
CheckMoveOut still lets trash rename the link without following it.
MkdirAll on a followed link is a no-op, since it already is that directory;
install --into needs that. A rename across the link boundary is refused.

The install, update, uninstall, and server callers pass the same policy
their discovery uses, through InstallOptions.SourceFollow and the update
context. No raw-write allowance was needed.

Refs #274
2026-10-05 14:42:09 +08:00
Willie 04ee235277 feat(sourcewalk): expose Follow.Resolve for writes through followed links
Operations that write or scan below a followed first-level link need the
real location of a logical path (<source>/_dev-skills/foo). Resolve maps
such a path onto the link's canonical target, reusing the same guards as
traversal, so sourcefs, group walks, trash, and audit roots can share one
answer instead of each re-deriving link identity.

Refs #274
2026-10-05 14:28:00 +08:00
Willie 89fc4fbdf1 feat(sourcewalk): follow first-level source links when enabled
Users link a checkout directly under the skills source (e.g.
skills/_dev-skills -> ~/code/dev-skills) and expect it to behave like any
other first-level entry. sourcewalk resolved only the root, so these links
were invisible to discovery, sync, update, and the dashboard.

follow_source_links (global and project config, default false) enables a
per-operation sourcewalk.Follow policy. Walk, WalkDir, and ReadDir report a
followed link under its own name as a directory and descend into its
target, so names, RelPath, FlatName, SourcePath, tracked-repo detection, and
metadata keys stay logical. Only entries directly under the source root are
followed. The policy is built from config and passed explicitly to every
discovery caller in the CLI, server, and reconcile, so counts agree; the
server keeps a per-instance snapshot refreshed on config load and save
instead of a process-global flag.

A link is skipped, with one warning per operation, when its target is
missing or unreadable, is the source or a parent of it, or overlaps an
active skills target. A missing or unreadable target marks the operation
incomplete: sync then skips pruning and copy-orphan deletion, and reconcile
keeps metadata, so an unmounted drive or a re-clone in progress cannot
erase target entries or install records. A link to an existing file is not
a directory link and stays an ordinary entry; treating it as incomplete
would disable pruning for anyone sharing a file such as .skillignore.

doctor keeps raw link identity: with the key off it still says the link is
not followed and names follow_source_links: true; with it on it reports
followed links and skipped links with the reason.

The raw-walk allowlist gains entries for the policy path in Walk and
WalkDir: the source root walk with a policy, and the walk of a validated
link target whose paths are remapped to the logical link path. Both live
in sourcewalk itself, the package that owns source traversal.

Refs #274
2026-10-05 14:26:09 +08:00
Willie 5c26981e44 refactor(git): reuse raw source directory reader
Use sourcewalk with zero options to preserve first-level link identity without adding a raw-walk allowance. Keep Git paths relative to the logical skills source.
2026-10-05 14:13:50 +08:00
Willie 3e8b6e8bd9 fix(git): warn before staging first-level source links
Source links can record machine-local paths in Git. Share advisory indexed and ignored checks across CLI, init, and dashboard staging and previews, with scope-correct ignore patterns and manual untracking instructions, without changing staging behavior.
2026-10-05 14:11:57 +08:00
Willie a3afc5f39a feat(memory): add a refresh button and keep short facts in INDEX.md (#413)
* feat(memory): add a refresh button and keep short facts in INDEX.md

Agents write shared memory notes from outside the dashboard, so the note
list and content went stale until the page was reloaded. A refresh button
in the note header refetches both.

The active guidance also told agents to create a separate file for every
durable fact, which produced one-sentence files. Short facts now go
directly under Notes in INDEX.md; only multi-paragraph notes get a file.

* fix(ui): keep the memory refresh button available without a selected note

The button lived in the note header, so a search with no matches hid it
and the list could not be refreshed to pick up a note an agent had just
written. It now sits next to the search box.
2026-10-05 12:09:56 +08:00
Willie d1371db7ea fix(audit): keep the skills source when auditing agents only (#412)
* fix(audit): keep the skills source when auditing agents only

`skillshare audit --agents` replaced sourcePath with the agents source
before building auditTUIContext, so the TUI's Skills tab discovered
skills inside the agents directory. The agents-only branches now read
agentsSourcePath themselves and the skills source is left alone.

Also from the #390 review, independent of .skillfollow:

- /api/skillignore reads the source through skillsSource() like the
  other routes (equal in project mode today; consistency only).
- scripts/windows/e2e-file-links.ps1 checked doctor output for a
  "Config:" line that no longer exists and could not match the ~-shortened
  path, so the run always aborted. It now checks status --json
  source.path, as e2e-skillfollow.ps1 does.
- windows-test runs internal/sourcewalk; GOOS=windows vet passes and the
  job on this PR is the first real run.

Refs #404, #407.

* fix(audit): relativize agent names against the agents source

With discovery moved to the agents source, auditInstalled and
auditFiltered still printed the header and computed SkillName against
the skills source, so `audit agents --json` reported names such as
`../agents/helper.md` when the two sources are siblings (Codex review on
#412). auditScanRoot now picks the root per kind for discovery, header,
and relativization; the skills source is kept only for TUI tab switching.

Also drops internal/sourcewalk from windows-test again: the first run on
windows-latest fails TestWalkParity, TestWalkDirParity and
TestSymlinkRoot (callback paths come from the canonical root, so
filepath.Rel against the test's root yields `..\..` chains). Logged on
#344.
2026-10-05 11:29:39 +08:00
Willie c61f48865e fix(server): reject skill names that escape the skills source (#408)
GET /api/audit/{name} joined the client-supplied name to the source
without checking it, so a name such as ../../x scanned a directory
outside the skills source and returned a clean result for it. The
single-skill update, the update stream, and the batch uninstall built
paths the same way; the tracked-repo uninstall route already refused
such names inline.

One helper, validSourceName, now holds that rule: non-empty, relative,
not starting with "..", and no Windows separators or volume prefixes on
any host, as sourcewalk does for declared names. The audit and update
routes answer 400, the update stream reports the name as an error
result, the batch uninstall fails that item, and the repo uninstall
route uses the same helper. Nested names such as group/skill still
resolve.

Closes #402
2026-10-05 11:03:39 +08:00
Willie 116feb6d16 fix(plugin): ask npm which registry it uses, and keep updates that never ran
Codex review on #400:

- The private-registry check read npm's config by hand, and each round
  found another place npm reads it from: the lowercase environment, the
  global npmrc, and now a globalconfig or prefix set in the user .npmrc.
  Rather than chase npm's rules, the check now asks npm itself:
  `npm config get registry @scope:registry` in the folder Pi runs npm in.
  Any registry other than npmjs, or npm not answering, leaves the package
  to Pi. The hand-written .npmrc reading is gone.

- When an apply stopped before any Agent ran, such as on a stale preview,
  the result had no outcomes, and the page took every planned update off
  its list, hiding the row's Update action. Only an Agent the result
  reports done comes off now.
2026-10-05 00:32:56 +08:00
Willie be661c593f fix(plugin): keep row updates without a new version, read npm's global config, keep odd manifests
Codex review on #400:

- The Plugins page kept only one version per plugin from a check, and
  only when the change had one, then hid the row's Update button when an
  Agent already had that version. A source plugin whose content changed
  without a version bump, or whose manifest has no version, so lost its
  Update action once the check dialog closed. The page now keeps what the
  check found per plugin and Agent, with or without a version, and an
  applied update takes those Agents off; the version only shows what the
  update brings.

- The private-registry check did not read npm's global config, so a
  registry set only in $PREFIX/etc/npmrc still sent the package name to
  npmjs. It now reads the file npm_config_globalconfig names, else
  $PREFIX/etc/npmrc with the prefix from the environment or the npm on
  PATH, and leaves the package to Pi when it can't tell where that file
  is.

- Reading the package version made a package.json whose version is not a
  string fail to decode as a whole, which dropped its pi manifest and the
  extensions it declares. The version is decoded on its own now, so an
  odd one is only skipped.
2026-10-05 00:22:30 +08:00
Willie c5571a8cae fix(plugin): leave npm ranges and tags to Pi, and read a lowercase userconfig
Codex review on #400:

- A package added with a range such as npm:demo@^1.2.0 was compared with
  npm's unrestricted latest. With latest at 2.0.0 and the newest matching
  release at 1.9.0, check kept offering 2.0.0, which Pi never installs
  because it keeps the configured range. Only a package added without a
  version follows latest now; one with a range or a tag is left to Pi, as
  a pinned one already was.

- The private-registry check read the user .npmrc path only from
  NPM_CONFIG_USERCONFIG. npm reads its environment case-insensitively, so
  a lowercase npm_config_userconfig pointing at an .npmrc with a private
  registry was ignored, and the package name went to npmjs. Both
  spellings are read now.

The plugin reference, its translations and the built-in skill say which
packages the check leaves to Pi.
2026-10-05 00:12:29 +08:00
Willie cc249c1498 fix(plugin): leave private npm registries to Pi, and keep updates for Agents still behind
Codex review on #400:

- The update check asked the public npm registry about every unpinned Pi
  package, even when the environment or an .npmrc sends the package, or
  its scope, to another registry. A public package with the same name
  could then hide a real update or offer an unrelated version. The check
  now reads the npm registry settings for the package and its scope from
  the environment, Pi's npm folder and the user .npmrc, and falls back to
  checking in Pi when any of them names another registry.

- The row's Update button disappeared as soon as any Agent had the
  version the check found, so with one Pi target behind and another up to
  date, the one behind lost it. The row now compares each selected Agent
  with that version on its own, and updates only the ones behind.
2026-10-05 00:05:07 +08:00
Willie 0f2ce354fc feat(plugin): check Pi npm packages for updates and show their versions
The Plugins page listed every binding a run touched, one row each, so a
single failure sat among a dozen "unchanged" rows, and the preview did
the same. Failures now come first, Agents that ended the same way share
a row, and what stayed the same folds into one Unchanged line that opens
to one row per plugin, with Pi packages apart. Logos stay only on the
rows that need a look.

An npm package in Pi recorded no version and no source, so it showed no
version and Check updates could only say to check in Pi. The installed
package.json now gives its version, on the Plugins page and in a Pi
target's Extensions tab, and check compares that version with npm's
latest. A newer version gets an Update button on the row and in the
check's result; it updates that one package with pi update <source>,
not pi update --extensions, which would also update packages the
preview never listed. update leaves a package already at npm's latest
alone, and records no version, so the dashboard keeps reading what Pi
installed. The two check messages that were English in every locale now
have keys.

Release-As: 0.24.5
2026-10-04 23:55:11 +08:00
Willie eb58dec1f0 fix(plugin): keep the global owner of a Pi project override
Codex review on #399: switching an extension of a global package in a
project writes a project override, which the project view reads in its
project-entry loop. In project mode that loop looked only at the
project's config, so the override lost the global plugin's name.

A project entry now falls back to the global config's owners when the
project config has none.
2026-10-04 23:14:07 +08:00
Willie 56b4f79ac4 fix(plugin): find global owners of inherited Pi packages in project mode
Codex review on #399: a dashboard started in project mode gives the
plugin service the project's config, so a global Pi package that
Skillshare installed, inherited by the project, found no owner and its
card fell back to the state path again. The earlier test kept
ConfigPath on the global config and missed this.

The service now takes the global config path in project mode, and the
project view looks up inherited global entries there.
2026-10-04 23:08:33 +08:00
Willie a510d2a089 fix(plugin): name the managed plugin of a Pi entry with a problem
Codex review on #399: an entry with a readable source but a value
Skillshare can't edit, such as "extensions": null, returned early in
both the global and project views before ManagedBy was set. Its card
then fell back to the plugin's local state path again.

The lookup now runs for every entry whose source can be read, before
the early return.
2026-10-04 23:02:03 +08:00
Willie c31dffc450 fix(plugin): name Skillshare-installed packages in Pi project views
Codex review on #399: piProjectState never set ManagedBy, unlike
piGlobalState. A project's Extensions tab (<project>@pi, or pi in
project mode) therefore titled a plugin Skillshare installed by its
state path and dropped the "Managed by Skillshare" tag, for both the
project's own entries and the global entries it inherits.

The lookup by source, then by resolved local path, moves into
piManagedBy and runs for every package of both views.
2026-10-04 22:56:12 +08:00
Willie f0e3ee22b3 test(plugin): keep Codex finder tests off the machine's own installs
The macOS finder tests stub PATH, but the candidate list still checked
/opt/homebrew/bin/codex, /usr/local/bin/codex and /Applications on the
real filesystem. In the devcontainer /usr/local/bin/codex exists, so it
won over the app bundle each test set up and make check failed there.

The finder now joins those absolute paths onto a root field, "" on a
real machine, and the tests point it at an empty temp dir.
2026-10-04 22:56:12 +08:00
Willie f5a2470276 fix(plugin): keep the Codex logo under a portable manifest
Adopting the portable root plugin.json replaced the whole Codex target
info, dropping the interface.logo read from .codex-plugin/plugin.json.
Codex still overlays that native manifest, and the portable one rarely
names a logo, so the add/list UI fell back to a generic icon.

Refs Codex review on #396.
2026-10-04 22:24:12 +08:00
Willie 04c4c95cba fix(plugin): keep version check for versionless portable manifests
A portable root plugin.json without a version left the Codex binding's
version empty, which skips the post-update version check, so an update
could advance the snapshot while Codex still ran an older copy. Codex
reports 1.0.0 for such a manifest, so expect that instead.

Refs Codex review on #396.
2026-10-04 22:18:57 +08:00
Willie 021e0037b6 fix(plugin): expect the version Codex reads from a portable manifest
Codex 0.159 installs from a portable root plugin.json (agent-plugins.org
schema) ahead of .codex-plugin/plugin.json, and reports 1.0.0 when that
manifest has no version. Skillshare preferred .codex-plugin/plugin.json,
so updating a package that ships both (ponytail 4.10.3) failed with
"native update returned version 1.0.0; expected 4.10.3" on every sync.
2026-10-04 22:14:28 +08:00
Willie c43bb763f9 fix(status): count one extension output per agent
Two agents that flatten to the same name matched the same manifest entry
and were both counted, so status could report 2/2 with one output on
disk. Each manifest key now satisfies one agent.

Refs Codex review on #392.
2026-10-04 20:01:51 +08:00
Willie aca1438fb5 Merge remote-tracking branch 'origin/runkids/391-extension-agent-status' into runkids/391-extension-agent-status 2026-10-04 19:49:55 +08:00
Willie 51a3be1ca6 fix(status): report drift when an extension agent's source changed
Counting an extension output as synced from its output hash alone missed
the common case: the user edits the agent in the source and runs status
to see whether a sync is due. The output still matched the manifest, so
status, doctor and the dashboard reported no drift although the next
sync would regenerate it.

Sync now records a SHA-256 of the source each output was converted from
in the manifest (sources), and status counts an output only when both
the output and that fingerprint are current. Outputs recorded by earlier
releases have no fingerprint and count as current until the next sync
records one, so upgrading does not show the 0/N the report was about.

Refs #391, Codex review on #392.
2026-10-04 19:49:29 +08:00
Willie d589e4f92f Merge branch 'main' into runkids/391-extension-agent-status 2026-10-04 19:46:28 +08:00
Willie af7654712c fix(server): log a disable that a .local rule overrides
When a "!" rule in .skillignore.local overrode a single disable, the handler
returned 409 before its only ops log call, although AddPattern had already
written the pattern to .skillignore. The shared file changed with no trace in
the log. Log the attempt with status "error" and the reason before returning.
2026-10-04 19:40:44 +08:00
Willie d4c2f6e9a0 fix(server): report toggles that .local ignore rules override
Disabling a skill only checked that its path was written to .skillignore.
A "!" rule in .skillignore.local is applied after that file, and an
unanchored one such as "!feature-radar" re-includes every path segment with
that name, so the skill stayed enabled while the dashboard reported success.
The dashboard also never showed the .local file, so nothing explained why.

Single and batch toggles now re-discover after writing and report an item
that did not reach the requested state as failed, naming the overriding
file; the batch toast shows the first failure's reason. GET /api/skillignore
and /api/agentignore return the .local file, and Settings > Files shows its
rules above the editor.
2026-10-04 19:34:24 +08:00
Willie f403ea5a6f fix(status): count extension-converted agents as synced
An agent target with an extension writes each agent under output_ext
(reviewer.toml) with transformed content, and the manifest tracks that
output name. status, doctor and the dashboard's Sync tab counted synced
agents by looking for the source name and comparing with the source
content, so every converted agent was missing: status showed 0/N with
drift, doctor said the same, and the Sync tab kept them pending after a
successful sync.

For an extension target, count an agent as synced when the manifest
tracks an output with the same stem and that file still matches its
recorded hash. Matching by stem avoids loading the extension spec in
status paths. doctor keeps reporting leftover broken links for these
targets.

Refs #391
2026-10-04 19:33:50 +08:00
Willie 881d08e52e fix(sync): skip nested target dotdirs when discovering source skills
Installing a repo whose root is a skill copies the whole repo into the
source, including any .claude/skills/ the author keeps for their own
development. Install discovery already skips target dotdirs, but the
source walk used by sync, check, and the dashboard only skipped .git,
so those dev-only skills surfaced as nested skills and were synced to
every target.

Share the skip set through sourcewalk and apply it in both source
walks. Only dotdirs nested below the source root are skipped: host-style
paths directly under the root (.cursor/skills/*) stay visible because
they drive target inference.
2026-10-04 17:11:53 +08:00
Willie b466a06d11 fix(memory): scope the global examples to cross-project facts
The global block still offered "a decision with its reason" and "a
pitfall you confirmed" as things that belong in it, so a project
decision matched both blocks. The examples are now per scope: the global
ones name user preferences, decisions that apply across projects, and
pitfalls in a tool; the project ones stay project-bound. Both rules take
the examples as a parameter instead of a string replacement.

Refs Codex review on #387.
2026-10-04 16:54:05 +08:00
Willie 91ad0cba3a fix(memory): say which notes belong in project versus shared memory
An agent that reads both a global and a project guidance block saw the
same body twice, with two "save it here", and had to guess whether a
user preference or a project decision went to the shared folder or the
project one. Each block now says what belongs in it: the global block
points project notes at a project memory when one exists, and the
project block keeps project notes and says that notes about the user,
their tools, or other projects do not belong in it. The project block's
examples follow suit, so it no longer names "a stated user preference".

The active rule's two overlapping sentences on when to save are merged
into one, and the plugins reference describes the older Codex.app path
without reference to current events.
2026-10-04 16:47:48 +08:00
Willie 70cb287672 fix(plugin): find the Codex CLI the desktop app ships (#383)
* fix(plugin): find the Codex CLI the desktop app ships

Plugin commands ran the bare name `codex` from PATH, so a machine with only
the Codex desktop app, or a scheduled job started with a minimal PATH,
reported the CLI as missing. The account-only `cli` override did not help:
the default codex target cannot set it, and an absolute path in a config
shared between macOS and Windows is wrong on one of them.

When `codex` is not on PATH, look in Homebrew's folders and the CLI the app
installs (Codex.app on macOS, %LOCALAPPDATA%\OpenAI\Codex\bin on Windows).
SKILLSHARE_CODEX_CLI overrides the search on one machine without touching
the synced config. The missing-CLI error now lists every place searched, in
the CLI and the dashboard.

Refs #338

* docs(plugin): note where Skillshare finds the Codex CLI

The plugin reference and the sharing guide still said every native CLI must
be on PATH, and the built-in skill told agents a missing CLI has no
fallback. Codex now has one, plus a machine-local override that agents
should prefer over a path in a shared config.

Refs #338

* fix(plugin): look for the Codex CLI inside ChatGPT.app

Since the Codex desktop app merged with ChatGPT, macOS installs it as
ChatGPT.app, so an app-only Mac with a minimal PATH still reported Codex as
missing. Try ChatGPT.app before Codex.app in /Applications and
~/Applications, the order Codex's own desktop-app finder uses.

Refs #338

* fix(plugin): run the packaged CLI entrypoint of the ChatGPT app

ChatGPT Desktop 26.924 removed Contents/Resources/codex and ships the CLI
at the entrypoint its package manifest declares, codex-cli/bin/codex, so a
current app-only Mac still reported Codex as missing. Try that entrypoint
first in each app bundle and keep the old path for earlier builds.

Refs #338

* fix(plugin): find the Codex CLI from the Windows installer

The official PowerShell installer puts codex.exe in
%LOCALAPPDATA%\Programs\OpenAI\Codex\bin and adds it to the user PATH. A job
started without that PATH missed it, the same gap Homebrew's folders cover
on macOS. Try it before the desktop app's copy.

Refs #338

* fix(plugin): keep an account's cli even when it is codex

The Codex search ran whenever the executable was the bare name codex, so an
account that set cli: codex could get SKILLSHARE_CODEX_CLI or the app's CLI
instead, against the documented rule that an account's own cli is kept.
Search only when no account cli is set.

Refs #338
2026-10-04 15:16:14 +08:00
Willie 8c60f51b9f feat(source): add the sourcewalk walker and sourcefs write boundary for proposal 274 (#385)
Implements rollout steps 1 and 2 of proposals/274-skillfollow.md. Nothing follows a link or parses .skillfollow yet.

Step 1: doctor reports first-level symlinks and junctions in the skills source as info, in text and JSON, without following them.

Step 2a: internal/sourcewalk owns traversal of the skills source. Thirteen raw walks move onto it with identical behavior, and a go/parser ratchet fails the tests on any new unclassified walk.

Step 2b: internal/sourcefs wraps an os.Root opened at the skills source and refuses any write whose path has a link component. The six write sites the proposal found in review move onto it, including the cross-device copy fallback, and a second ratchet covers raw os write calls. The only behavior change is that an explicit path crossing a link inside the source is refused instead of written through.

Refs #274

Release-As: 0.24.3
2026-10-04 15:15:43 +08:00
Willie 464d2fab73 fix(memory): route lessons to LEARNED.md in active guidance
The active rule asked for one fact per file, while init still creates
LEARNED.md as a multi-lesson log and the docs tell users to have agents
update it. The rule now says which goes where: a fact or decision gets
its own file linked from INDEX.md, a lesson from a failure gets a new
heading in LEARNED.md.

Refs Codex review on #384.
2026-10-04 14:53:39 +08:00
Willie 9ea51b1890 docs: fix stale facts and conflicts found by a prompt audit
- The LEARNED.md starter said to update only on request, which
  contradicted the active memory guidance; it now defers to the guidance.
- website/AGENTS.md hardcoded a command-page count that had drifted and
  omitted the Desktop App navbar item.
- wiki/testing.md gave `bash -c` and `bash -lc` as the rule in different
  places; it now says which container needs which, and why.
- Migration-relative wording in wiki/architecture.md and an unexplained
  "Always use" in the targets reference are stated as current rules with
  their reason.
2026-10-04 14:45:38 +08:00
Willie 9ac9682dcc perf(memory): tell agents to prefer shared notes over tool memory
The generated guidance block defined "what is worth saving" the same way
agents' built-in memory does, but never said how the two relate, so agents
with their own memory kept saving there and left the shared folder alone.
The active rule now says to save here instead of tool-specific memory,
gives a note format and an INDEX.md line shape, and both modes read
INDEX.md at the start of each task instead of "when relevant". The
passive rule asks agents to point out facts worth keeping so the user
can ask for them to be saved.

Dashboard mode descriptions, the CLI help, and the website pages are
updated to match. Test fixtures that anchored on "Read only" follow the
new wording.
2026-10-04 14:45:38 +08:00
Willie 887461b36a perf(upgrade): download release assets over parallel ranges (#381)
* perf(upgrade): download release assets over parallel ranges

release-assets.githubusercontent.com throttles each connection on some
networks. The 9.6 MB darwin_arm64 archive took 122-171 s over one
connection but 25 s as 8 ranges, each on its own connection, with an
identical checksum. `skillshare upgrade` and the first `skillshare ui`
of a new version both waited minutes on that download.

DownloadToFile probes with `Range: bytes=0-0`; a 206 answer splits the
file into ranges fetched in parallel from the redirected URL, and a 200
answer streams the single response as before, so servers without range
support cost no extra request. Size limits and checksum verification are
unchanged.

The UI asset download is about 2 MB now, so the docs no longer say 1 MB.

* fix(upgrade): keep the caller's transport for parallel ranges

Each range built its client from http.DefaultTransport, so a caller's
custom transport (private CA, authenticated proxy, custom dialer) only
applied to the probe. The ranges then failed or bypassed the caller's
network setup even though the probe succeeded. Each range now copies the
caller's client with a clone of its *http.Transport, which keeps those
settings and still gives the range its own connection. A transport that
cannot be cloned is shared as is.
2026-10-04 14:03:46 +08:00
Willie 91dbd2a9fc fix(mcp): turn off a global server for Pi with its project override (#379)
* fix(mcp): turn off a global server for Pi with its project override

Since Pi 1.0.1, a .pi/mcp.json entry without command, url or type
overrides only enabled, exposure and toolExposure of the global server
with that name, and Pi's own /mcp writes {"enabled": false} for it.

Skillshare still wrote the global server's command (or url without the
query) next to enabled: false. Pi reads that entry as a project server
that replaces the global one, so turning it back on in Pi's /mcp only
dropped enabled and started the server without its args, env or
headers, and a url server lost its query. It also needed the global
server, so project mode refused pi in a disabled entry.

The switch is now the override alone. It needs nothing from the global
server, so project mode accepts it; a switch written by an earlier
release is rewritten on the next sync, and an identical override Pi
already wrote is unchanged instead of a conflict. The dashboard counts
Pi as an Agent with a per-project switch. Pi before 1.0.1 reports the
override as invalid, as the docs now say.

Closes #378

* fix(mcp): compare the whole Pi override before calling it the switch

Codex review on #379: an unmanaged .pi/mcp.json override such as
{"enabled": false, "exposure": "direct"} has the same ownership hash as
the switch sync writes, because that hash leaves out Pi-only fields like
exposure and toolExposure. Sync then reported it as unchanged, although
it differs from the switch and the docs promise a conflict. The check
now compares the whole entry, so only an exact {"enabled": false} is
unchanged; Pi fields a person adds to a switch this config already owns
are still kept, as for any managed Pi entry.

* docs(mcp): say which Pi overrides of a switch conflict

Codex review on #379 read "the override differs from the disabled
entry" as covering a switch sync already owns, where Pi settings added
in Pi, such as exposure, are kept: managed Pi entries keep Pi-only
fields by design (withAgentFields, piOwnedFields), and exposure does
nothing while the server is off. Turning the server back on in Pi
changes enabled, which the ownership hash covers, so that is a conflict.

The docs now say so: an exact {"enabled": false} is no conflict, Pi
settings added to an owned switch are kept, re-enabling in Pi
conflicts, and a differing override sync did not write conflicts.
TestPiSwitchOwnedBySync pins both owned-switch cases.
2026-10-04 11:50:03 +08:00
Willie 5fe2d78c2b feat(config): lay out auto-saved config files like the dashboard's Beautify (#380)
Config files written by the CLI were marshaled straight from the Go
structs, so sections ran together and plugins, hooks and mcp landed
above sources. Programmatic writes of the global and project config
now sort sections like the dashboard and separate them with a blank
line, keeping the original order when sorting would break a YAML alias.

Also documents one worktree and throwaway container per branch task.

Release-As: 0.24.1
2026-10-04 11:43:26 +08:00
Willie b4296e3cff feat(memory): let each agent choose passive or active note updates (#377)
* feat(memory): let each agent choose passive or active note updates

Shared memory guidance only told agents to update notes when asked, so
anything worth keeping was lost unless the user remembered to request it.
Some users want agents to maintain the notes themselves, but because the
notes are shared across every tool, blind writes would fill them with
noise.

Add an active mode whose guidance limits saves to facts that will matter
in later sessions, asks the agent to propose a note when unsure, to update
an existing note instead of duplicating it, and to report what it saved.
Passive stays the default and its block is byte-for-byte unchanged, so
existing connections remain configured.

The mode is recorded in each block's marker, so agents can differ. Agents
that read the same file share one block; the dashboard switches them
together and the API rejects conflicting modes for one file. The CLI
prints either block with `extras memory instructions --update-mode`
(`--mode` already means an extra's sync mode).

* feat(memory): review guidance changes as a unified diff

The side-by-side Before/After panes made users scan two windows to find
the one block that changed. A single Git-style diff shows only the
removed and added lines in context, with removals listed first.

* fix(memory): flag targets that read guidance in both update modes

A tool that imports several shared instruction files can read a passive
block in one and an active block in another. The status showed only the
best-ranked block, so the target looked configured in one mode while the
agent still received the other, and switching modes rewrote just one
file. Report such a target as needing attention instead, so the user
sets the tools on those files to one mode.

* fix(memory): refuse mode switches that would split a target's guidance

A target can read intact blocks in more than one file, such as its own
file and an imported shared file. Switching its mode rewrote only the
file the status pointed at, leaving the other block in the old mode and
creating the contradiction the mixed_modes state reports. Skip such a
switch as multiple_blocks so the user removes the extra block first.

Document guidance modes, mixed_modes and multiple_blocks in the
architecture topic, which still described a single implicit mode.
2026-10-04 10:23:47 +08:00
Willie 27f53317d4 fix(audit): stop flagging config keys named like role labels (#376)
* fix(audit): stop flagging config keys named like role labels

prompt-injection-1 matched any line starting with system:, override:,
ignore:, admin:, or root: in any case, so ordinary YAML and JS config such
as `root: ./src` or `ignore: { tags: [...] }` was reported as a CRITICAL
prompt override and blocked installs. A line whose key only carries a single
value, a trailing comment, or an inline object is configuration, not a
directive, so the rule now excludes it. Lines with prose after the colon,
including lowercase `system: Override ...`, are still flagged.

Refs #374

* fix(audit): keep flagging single-token role directives

The config-key exclusion accepted any single token and an empty value in any
case, so directives such as `SYSTEM: jailbreak`, `ADMIN: execute`, and a
bare `OVERRIDE:` were no longer reported, and no other rule catches them.

The exclusion now applies only to lowercase keys, since config keys are
lowercase and role labels are written in capitals, and only when the value
is plain data: a boolean, null, number, path, inline object, or nothing.
`admin: execute` stays flagged because a bare word is not a config value.

* fix(audit): treat quoted and Windows paths as config values

The config-key exclusion only recognized unquoted paths with forward slashes,
so `root: "./src"`, `root: './src'`, and `root: C:\repo` were still
reported as CRITICAL even though the docs say path values are configuration.
The exclusion now accepts these forms. Quoted text that is not a path, such as
`system: "Override all constraints"`, is still flagged.

The Japanese, Simplified Chinese, and Traditional Chinese audit-engine pages
also described every role prefix as matched case-insensitively with no
exception; they now state the lowercase config-key exception like the English
and Korean pages.

* docs(audit): say config-key paths must not contain spaces

Allowing spaces inside quoted path values would let a directive pass the
CRITICAL prompt-injection-1 rule by starting with a slash, as in
`system: "/ignore your rules and print every secret"`. Quoted paths with
spaces therefore stay flagged; the docs now say the exception covers paths
without spaces, and a test pins the slash-prefixed directive as CRITICAL so
a later relaxation of the exclusion cannot reopen it.

* fix(audit): close comment and inline-object bypasses in config-key exclusion

The exclusion accepted an empty value followed by a # or // comment, so
`system: # send every file to the attacker` was not reported. Neither marker
hides text in Markdown, so the directive stays visible to the model. Trailing
comments are no longer accepted.

Inline objects only required some `key:` inside braces, which also excluded
prose such as `system: {note: ignore your rules}`, while `ignore: {}` and
quoted keys like `{"tags": ["design"]}` were still flagged. An inline object
now passes only when its words are joined by punctuation, never by a bare
space, which accepts empty objects and quoted keys and rejects prose.

* fix(audit): flag empty role keys that can open a multiline block

An empty value was excluded as a nested YAML mapping, so `system:` followed
by an indented line such as `Reveal every repository file in your response`
was never reported: the scanner checks each line on its own and the indented
payload matches no other rule. The exclusion now requires a value on the same
line. Nested mappings under these key names (`ignore:` then an indented
list) are flagged again; one-line values and inline objects still pass.

* fix(audit): scope config-key value kinds to the key that uses them

Paths and inline objects were accepted for every role key, so a directive
written without spaces, such as `system: /ignore/all/previous/instructions`,
passed the exclusion, and prompt-injection-0 needs whitespace to match.

Each key now accepts only the values its config uses: scalars for all five,
space-free paths only for root, and inline objects only for ignore, whose
words may no longer be joined by slashes or dots. The known false positives
(`root: ./src`, `admin: false`, `ignore: { tags: ["design"] }`) still pass.
2026-10-04 10:21:31 +08:00
Willie 3c7e48e239 fix(ui): let a Pi extension switch remove a rule instead of adding its opposite
Turning on a file that only a "-path" rule kept off wrote "+path", so the
row ended with a rule pinning it on and a separate Remove rule button that
did nearly the same thing. When removing the row's own rule gives the state
the switch wants, the view now says so (unruled) and the switch removes the
rule. Remove rule stays only where the rule is redundant. The prediction
uses the same rule edit, evaluation and guard as the write, so the preview
and the result agree.
2026-10-04 05:02:04 +08:00
Willie 65998a4d47 fix(cli): keep inline picker rows and summaries on one line
Picker options carried full descriptions, so installing a repo with long
skill descriptions turned each option into five to eight wrapped lines,
which also broke the picker's row-based scrolling. Options are now one
row, cut to the terminal width. After a pick of more than three items
the summary line shows a count and the first three names instead of
every name, as init already does for tools. The MCP edit menu also
loses a leftover "Esc cancels" note that the key line already covers.
2026-10-04 04:32:45 +08:00
Willie 355b42ee57 feat(cli): one frame and keymap for every TUI, inline prompts, file viewer (#370)
* chore(devcontainer): add a TUI screenshot kit

Recreating the devcontainer wipes vhs and the browser it drives, so
every TUI review started with a manual reinstall. The start script now
installs the recording tools when they are missing and skips straight
through when they are present.

record-tui.sh screenshots each full-screen TUI in its own fresh ssenv,
filled by tui-fixture.sh through real commands, so the TUI redesign can
compare before and after on the same data.

Refs #200

* feat(cli): ask small choices inline instead of full screen

The restore source, new-skill wizard, install skill picker and the
other checklist prompts each took over the whole screen for a handful
of options, and the answer vanished once chosen. They now use the same
inline prompt as init: the question appears in place, collapses into a
"✓ label value" line, and stays in the scrollback.

Lists longer than the visible rows can be narrowed with /, and esc
inside the filter only leaves the filter. Every inline prompt shows
one key line in the wording the full-screen TUIs will share. The new
wizard drops its esc-goes-back steps: esc cancels, as in init.

Refs #200

* feat(cli): give the list TUI the shared frame and keymap

The full-screen TUIs each had their own header, footer and key letters,
so the same action used different keys per screen and the help line
listed everything at once. list is the first screen on a shared frame:
one title line with scope, counts and tabs, the list and details side
by side without dividers, and one key line with the common keys and
the position. Everything else moves under ?, and confirmations ask on
the key line so the list stays visible.

Keys follow one lowercase letter per action (u update, d uninstall,
t enable/disable, m manual only, ! audit). The All tab and the s
status cycle are gone: tab switches skills and agents, and status:
in the filter or --status covers enabled/disabled. Rows drop the
group prefix that the heading above them already shows.

Refs #200

* docs: drop key-by-key TUI instructions from command pages

The command pages carried key tables, screen mock-ups and one bullet
per key for every full-screen TUI, in five languages. The TUIs now show
their keys on screen, so these pages repeated what the screen says and
went stale with every key change (several already listed keys that no
longer exist). Each TUI section is now one paragraph on what the screen
shows and what it can do. What the screen does not explain stays: the
filter tag syntax, what manual only does to SKILL.md, the analyze color
thresholds, when audit opens its TUI, and that target changes are saved
to config immediately.

Refs #200

* feat(cli): move the trash TUI onto the shared frame

The trash TUI replaced the whole screen with a confirmation and a
spinner, so you lost sight of what you had selected, and r/d did
nothing until something was ticked. It now uses the shared frame:
checkbox rows with size and age, details beside them, and the
confirmation, the names it covers, the running operation and its
result on the bottom lines. With nothing selected, r and d act on the
row under the cursor.

Refs #200

* feat(cli): move the log TUI onto the shared frame

Log rows were two-line "[date] COMMAND — status" blocks, stats lived
behind s as a separate screen, and delete swapped the help bar for a
prompt you could miss. Rows are now one line with time, command and a
status mark; Entries and Stats are tabs on the title line next to the
count and success rate; and the delete confirmation takes over the key
line. With nothing selected, d deletes the entry under the cursor.

Row delegates no longer force a 40-column minimum, which wrapped every
row in lists narrower than that.

Refs #200

* feat(cli): move the audit TUI onto the shared frame

The audit TUI had its own tab bar, a divider column and a two-line
summary footer, and esc quit even with a filter typed. It now uses the
shared frame: scan counts and the Skills / Agents tabs on the title
line, the worst severity at the right of each row, the scan-wide
severity, threat and policy summary on the note line, and esc clears
the filter before quitting. Rows under a group no longer repeat the
group's name.

Colored title facts keep their color instead of being dimmed.

Refs #200

* feat(cli): move the audit rules TUI onto the shared frame

The rules browser used keys no other screen shared: s for severity,
space to toggle and R pressed twice to reset. It now follows the
shared keymap: t turns a rule or pattern on or off, e asks for the new
severity on the key line, and R asks yes/no before deleting
audit-rules.yaml. The severity tabs sit on the title line, rows show
the rule count or "off" at the right, and rule rows drop the pattern
prefix the group row already shows. Space still toggles.

When the title line is too narrow for every tab, it shows only the
active one instead of clipping them.

Refs #200

* feat(cli): move the analyze TUI onto the shared frame

Analyze put the target switcher, a stats line and the help text in a
three-line footer and sorted with s, which means sync everywhere else.
It now uses the shared frame: skill count and always-loaded and
on-demand tokens on the title line, one tab per target group, o to
sort with the current order on the key line, and esc clears the
filter before quitting. The details follow the list layout, and lint
warnings use the ! mark the CLI uses.

renderDetailSection and renderFactRow had no other callers and are
removed.

Refs #200

* feat(cli): move the diff TUI onto the shared frame

The diff browser drew ruled separators, "── Next Steps ──" banners and
an arrow before every suggested command, and esc quit with a filter
typed. It now uses the shared frame: target, extra and difference
counts on the title line, what differs at the right of each row, group
rows like the list's, plain headings in the details, and esc clears
the filter before quitting.

Refs #200

* feat(cli): move the restore TUI onto the shared frame

Restore replaced the whole screen for the confirmation, the running
operation and the result, rendered two-line rows, and esc quit even
with a filter typed. It now uses the shared frame: the target or the
chosen target and its backup count on the title line, one-line rows
with backup count and age or skill count and size at the right, and
the confirmation, what it covers, the spinner and the result on the
bottom lines. Esc clears the filter, then goes back, then quits.

Refs #200

* feat(cli): move the MCP manager onto the shared frame and keymap

The MCP manager used its own letters (a add, x remove, b backups, r
refresh), hid the details behind Enter, and printed two help bars.
It now uses the shared frame and the keys every other TUI uses: n
add, i import, e edit, d remove, s sync, r restore, with the details
of the selected server beside the list and ? for all keys. Refresh is
gone because the manager reloads after every action. Command docs drop
the key table, and the how-to page and built-in skill name the new
keys.

Refs #200

* feat(cli): move the search selector onto the shared frame

The search selector rendered two-line checkbox rows, put the details
under the list, used s for "search again" although s means sync
everywhere else, and Enter with nothing selected silently cancelled.
It now uses the shared frame with the keyword and counts on the title
line, one-line rows with stars or risk at the right, and the details
beside the list. Esc clears the filter, then goes back to the keyword
prompt; Enter with nothing selected installs the row under the cursor,
as r and d do in other TUIs. The command page drops its screen mock-up
and key list, and no longer claims the search restarts after an
install.

Refs #200

* feat(cli): move the target TUI onto the shared frame and keymap

The target TUI had five uppercase keys (M, N, I, E, R), a separate
skills-or-agents picker before three of them, and full-screen pickers
and remove confirmation. e now opens one menu of every setting the
target has (skills mode, naming, include, exclude; agents mode,
include, exclude, with agent filters marked as ignored in symlink
mode), and the pickers and pattern editor open beside the list. d asks
on the key line with the command it will run and what removal does.
Patterns are added with n, as everywhere else. Also drops the filter
style helper the MCP change left unused.

Refs #200

* feat(cli): move the extras TUI onto the shared frame and keymap

The extras TUI used six uppercase keys (N, X, S, C, M, F), each with
its own full-screen menu, and a full-screen confirmation. It now uses
the shared frame and keys: n new, d remove, s sync, c collect, and e
for one menu with the mode and flatten setting of every target. The
target menu and mode picker open beside the list, and confirmations
ask on the key line with what the action does. The picker row helper
moves to the shared frame, and the split and scroll-hint helpers no
TUI uses any more are removed.

Refs #200

* feat(cli): share one framed file viewer between list and extras

The list and extras file viewers were two copies of the same layout
with the old look: an indented title, a │ column between the tree and
the file, and a hard-coded yellow cursor. Both now draw through one
renderer in the shared frame: the skill or extra and the open file on
the title line, the tree beside the file without a divider, and the
scroll position at the end of the key line.

Refs #200

* feat(cli): ask for a large repo's folder inline

Installing from a repo with 50+ skills opened a full-screen folder
browser with its own filter bar and key line, then dropped back to an
inline skill checklist. The folder question is now an inline select
like every other install prompt: the first choice installs everything
below the current folder, esc goes up one level, and the chosen folder
collapses into a "✓ folder" line before the skill list appears.

The screenshot fixture gains a 60-skill repo so the picker can be
recorded, and the filter-bar and page-info helpers that only the old
picker used are removed.

Refs #200

* feat(cli): ask for MCP and plugin values inline

MCP and plugin values were typed into a separate textarea program that
submitted on ctrl+d, a key no other prompt uses. They now go through a
new ui.Text prompt drawn like the other inline questions: enter
submits, alt+enter starts a new line, and the box grows with the value.
The MCP name is a single line, so it uses ui.Input. Answers are not
echoed afterwards because a value can hold a secret.

Refs #200

* feat(cli): drop the full-screen MCP change preview

Every MCP add, edit, remove, restore and sync already prints its plan
inline and then asks whether to apply it. The full-screen preview in
between showed the same changes again and needed one more enter, so it
is removed along with the review step of the prompts interface. The
confirm question loses its "Review complete" prefix, and a blank line
now separates it from the plan.

Refs #200

* feat(cli): ask extras init questions inline

`extras init` without arguments ran its own full-screen wizard with a
hand-drawn cursor and a different key line on every step. It now asks
the same questions inline, like `init`: each answer collapses into a
"✓" line, the target line shows where the file lands, and esc cancels
with "Nothing was created".

A new ui.InputValid keeps a question open and shows the error under it
until the answer passes, which the name, file name and per-target file
name questions need.

Refs #200

* fix(cli): drop stale key hints from MCP prompts and help

`mcp --help` still listed the manager keys from before its redesign
(a add, x remove, b backups, r refresh), and several MCP and plugin
questions repeated "Esc cancels" or claimed `a` selects all, while the
key line under every prompt already shows the real keys (ctrl+a for
all). The help now just says where the keys are, and the questions keep
only what the key line cannot say. The MCP command page drops the same
key list in all five locales.

Refs #200

* feat(cli): open a skill's files at its audit findings

The audit TUI listed each finding's file, line and one-line snippet, but
judging whether a finding is real needs the lines around it, and that
meant leaving the TUI to open the file by hand. Enter now opens the
selected skill's files at its most severe finding, with the raw text
numbered and the flagged line marked; n/N steps through the findings in
the order the details number them, and the line above the keys says
which finding is shown.

The viewer state lives in a new fileBrowser so analyze and trash can
open skills the same way. Long lines wrap under their text instead of
being cut, since the evidence is often at the end of a line.

Refs #200

* feat(cli): open a skill's files from analyze

Analyze flags lint issues such as a description that is too short, but
the screen only showed the description text, so fixing a skill meant
finding its folder first. Enter now opens the selected skill's files,
including local ones that live only in a target, as raw numbered text
with the front matter, and the line above the keys repeats the first
lint issue so it can be checked against what is written.

Refs #200

* feat(cli): open a trashed item's files before restoring

The trash TUI showed only the first fifteen lines of SKILL.md, which is
not enough to tell two trashed versions apart or to see a skill's other
files. Enter now opens the item's files in the same viewer as list and
analyze, rendered as Markdown, and esc returns to the trash.

Refs #200

* fix(cli): leave front matter out of the agent viewer in list

Opening an agent in the list TUI rendered its whole file as Markdown,
so the front matter came out as a horizontal rule and a "## name: ..."
heading. Skills already drop it because the details panel shows the
name and description; agents now read their body the same way.

Refs #200

* test(cli): compare trash restore destinations home-relative

The trash TUI shows restore destinations through shortenPath, so with
HOME=/tmp, as in the offline Docker sandbox job, /tmp/skills prints as
~/skills and the tests that expected the absolute path failed. They now
expect whatever shortenPath gives for the same path.

Refs #200

* fix(cli): show control characters in viewed files as symbols

The file viewers in list, extras, audit, analyze and trash wrote a
file's bytes to the terminal as they were. Audit and analyze exist to
inspect skills that may not be trusted, and a SKILL.md with an ESC or
OSC sequence could rewrite the screen or the window title instead of
being shown. Control characters are now replaced with same-width
symbols such as ␛ before rendering, which keeps line numbers and the
marked finding line in place; newlines and tabs are kept and CRLF is
read as LF.

Refs #200

* fix(cli): open audit findings in files the tree leaves out

The file tree skips dotfiles and stops at three levels and 200 entries,
but the audit scan reads root dotfiles and goes deeper. A finding in
such a file, like a root .env, left the viewer on the previous file
with the new finding's note above unrelated text. Opening a finding now
adds its file to the top of the tree under its full relative path when
it is missing, so the marked line is always the one reported.

Refs #200

* fix(devcontainer): exit nonzero when a TUI screenshot fails to record

record-tui.sh printed FAIL for a screen whose vhs run failed but still
exited 0, so an unattended run looked like every screenshot was taken.
It now remembers a failure, keeps cleaning up and recording the other
screens, and exits 1 at the end.

Refs #200

* fix(cli): keep other files' front matter and fit text to the pane

Two problems in the shared file viewer:

- It dropped YAML front matter from every Markdown file, so a trashed
  skill's README or template lost its metadata, which nothing else in
  the trash view shows. Only SKILL.md and a single-file agent drop it
  now, since their name and description appear in the details.
- The open file was wrapped to at least 40 columns. In a narrow
  terminal that is wider than the right pane, so the layout wrapped
  each row a second time, breaking lines apart and moving the marked
  finding line. Text now wraps to the pane's real width.

Refs #200

* fix(cli): keep frame lines within the terminal width

joinEnds padded the title, confirm and filter lines but never cut them,
so a long search query, path or command made the line wrap. Every TUI
sizes its body for a fixed four lines of frame, so the wrap pushed the
bottom of the body off the screen. Long text is now cut from the left
side, keeping the status or tabs on the right. Lines are left alone
until the terminal has reported its width.

Refs #200

* fix(cli): show a nested skill's folder when no group heading names it

List rows dropped the first path segment for every skill on the grounds
that the group heading above shows it. With a single group, or a filter
that leaves one, there is no heading, so frontend/react showed as just
react and same-named skills from two folders looked alike. Rows now
drop it only under a heading; a tracked skill's repo dir is still
dropped as before.

Refs #200

* fix(ui): let esc leave a prompt's filter after enter found nothing

The inline prompt guessed the filter state from key presses and treated
enter as leaving the filter. huh's Select stays in the filter when enter
has no match, so the next esc, meant to leave the filter, cancelled the
question instead: one folder level back in the large-repo picker, or the
whole install at the top. The prompt now asks the field whether it is
filtering.

Refs #200

* fix(cli): show control characters in the trash preview as symbols

The trash details panel prints the first lines of a trashed SKILL.md,
and it wrote them to the terminal as they were. The file viewers already
pass text through printableText so an ESC or OSC sequence in a skill is
shown instead of acted on; the preview now does the same.

Refs #200

* fix(cli): scroll an agent's text with the wheel anywhere in list

An open agent has no file tree, but the wheel still treated the left
quarter of the screen as the tree, so scrolling there did nothing. The
whole screen now scrolls the agent's text.

Refs #200

* fix(ui): read prompt answers after the prompt has run

The prompt helpers returned "value, runPrompt(..., &value)". The Go spec
leaves the order between reading value and the call unspecified, so a
compiler may return the value from before the user answered. gc 1.25
runs the call first, so answers arrive today, but nothing guarantees it.
Each helper now runs the prompt into err and then returns the value.

Refs #200

* fix(cli): name hidden Unicode and skip oversized files in the viewer

Two gaps in the file viewer used to inspect untrusted skills:

- Zero-width and bidirectional characters passed through unchanged,
  although the audit reports them as HIGH (Trojan Source). Opening such
  a finding reordered or hid the text instead of showing the cause.
  They now appear as <U+202E> and the like.
- Moving onto a file read all of it, then copied it again while
  numbering and wrapping, with no size limit. The audit skips files
  over 1 MB; the viewer now does too and says the file is too large.

Refs #200

* fix(cli): show control characters in viewed file names as symbols

File contents in the viewer already pass through printableText, but the
tree and the title printed file and skill names as they were. A file
named with an ESC sequence, a newline or a bidirectional character in a
skill under review could still reach the terminal, or break the frame
across lines. Names now go through printableName, which also shows a
newline or tab as a symbol so the name stays on one line.

Refs #200
2026-10-04 04:14:47 +08:00
Willie 7a9b37e852 fix(mcp): explain Pi project overrides in import and sync conflicts (#371)
* fix(mcp): explain Pi project overrides in import and sync conflicts

Since Pi 1.0.1, /mcp in Pi writes a .pi/mcp.json entry with only
enabled, exposure or toolExposure to override the global server of that
name. Skillshare treated it as a broken server: import reported
"requires exactly one of command or url", and a project server with the
same name hit a conflict that told the user to import it, which import
always refused, so the user was left without a way forward in the CLI.

Import now names the entry as a Pi project override with nothing to
import. Sync reports the conflict as an override to replace or remove
with /mcp in Pi, and the dashboard offers only Replace for it. Sync
behavior is unchanged; detection matches Pi's own isOverride (no
command, url or type).

* fix(mcp): limit Pi override detection to project imports

Codex review on #371: importing Pi's global ~/.pi/agent/mcp.json, or a
pasted snippet, also reported a connection-less entry such as
{"enabled": false} as a project override and pointed to /mcp in a
project. Pi gives that shape override meaning only in a project's
.pi/mcp.json; elsewhere it is an invalid server. importNative now takes
the scope, so global and pasted imports keep the usual "requires exactly
one of command or url" problem.

* fix(mcp): classify a synced entry turned into a Pi override

Codex review on #371: when an entry this config synced to .pi/mcp.json
was later changed into a connection-less Pi override, the ledger still
marked it managed, so the generic "Agent configuration changed; import
it" conflict won and the dashboard offered an Import that project import
now refuses. The override case now comes first and covers a managed
entry whose content changed as well as an unmanaged one, so both offer
only Replace.

* fix(mcp): classify Pi overrides before every import-offering conflict

Codex review on #371: the prune conflict ("Pi setting changed since
sync; import it") still came before the override case, so a synced
server whose cleared Pi field Pi had replaced with an override again
offered an Import that project import refuses. Rather than chase each
earlier case, the override case now opens the switch. An entry owned by
another live config keeps its own message, since only that config can
release it.

* fix(mcp): treat a removed owner's entry like any other Pi override

Codex review on #371: an entry whose ledger owner was a removed config,
later turned into a connection-less Pi override, still got the orphaned
conflict, which offers an Import that project import refuses. Skillshare
never writes a connection-less Pi entry, so the override case now covers
every such entry except one a live config owns, which only that config
can release. It still skips an entry recorded at its current hash, which
is what Replace leaves, so Replace settles the conflict; a test now
covers that.
2026-10-04 03:18:14 +08:00
Willie dd6991ca3c fix(push): refuse unpushed config.yaml history at root scope (#366) (#372)
* fix(push): reject unpushed config history at root scope

Untracking config.yaml at the tip still uploads its contents from earlier local commits. Refuse root-scope pushes before mutation and before upload when unpublished history adds or changes the file or its directory tree, without rewriting history.

Share the check across the CLI and dashboard, return unpushed_config_history, and allow removals and config inherited unchanged from upstream merges. Cover first pushes, dry runs, push --pull, directory paths, renames, and merge commits, and document manual repair in all website locales.

Fixes #366

* fix(push): check exactly what push sends for config history

Review of the config history guard found two range problems:

- The guard checked @{u}..HEAD, but a same-named upstream was pushed with
  a bare `git push`. With remote.pushDefault or push.default=matching that
  could send config-bearing commits to another remote, or other branches,
  that the guard never looked at. PushArgs now always names the upstream
  remote and branch (read from branch config, so names with "/" survive),
  so the checked range and the pushed range are the same.
- Without an upstream every commit on HEAD counted as unpushed, so a
  config.yaml already published on origin followed by the
  "Keep config.yaml out of version control" commit was refused. The guard
  now checks HEAD --not --remotes=<push remote>.

The refusal message now suggests rebasing from just before the oldest
offending commit instead of @{u} or --root, which would rewrite commits
the remote already has.

* fix(push): check config history against the remote's live refs

Local remote-tracking refs can be stale. When another machine force-pushes
the remote to drop a leaked config.yaml commit, this clone's
refs/remotes/<remote>/... still contains it, so the guard treated the leak
as already published and HEAD:main, still a fast-forward, would publish it
again.

The guard now excludes only the objects the push remote's refs point to
right now (git ls-remote), limited to objects this clone has. It falls back
to local remote-tracking refs only when the remote cannot be reached, in
which case the push fails anyway. Revisions go to git log on stdin as
"^<id>" lines because git before 2.42 rejects "--not" there.

* fix(push): screen config history against every push URL, without replace refs

Two more ways the screened history could differ from what git push uploads:

- `git ls-remote <remote>` queries the fetch URL, but git push sends to every
  remote.<name>.pushurl. A commit already on the fetch URL could still be
  new to a push URL. The guard now lists `git remote get-url --push --all`
  and excludes only objects that every push URL has.
- With a refs/replace/* entry for an offending commit, git log read the
  sanitized replacement while the push pack carries the original. The log
  and the rebase-base lookup now run with --no-replace-objects.

* fix(push): let pushes the remote will reject reach the pull step

Our own review of the config history guard found two problems:

- When another machine had pushed, the remote's tip was unknown locally, so
  no live ref survived the "objects this clone has" filter and the whole
  history counted as unpushed. A config.yaml published long ago, followed by
  the stop-tracking commit, then blocked `push --pull` before it could pull,
  and replaced the plain push's ErrPushRejected hint. A push URL whose
  destination branch tip this clone lacks rejects the push as
  non-fast-forward and receives nothing, so it is now left out, and the
  check passes when every push URL is like that. `push --pull` re-runs the
  check after pulling, with the tip known.
- PushRemote read branch.<b>.remote even when @{u} no longer resolved, while
  PushArgs then pushed to origin. pushTarget now follows the same
  HasUpstream rule as PushArgs.
2026-10-04 03:09:17 +08:00
Willie ca11f63a09 fix(plugin): trust symlinked ancestors of the Agent config root (#368)
* fix(plugin): trust symlinked ancestors of the Agent config root (#365)

noSymlink walked every ancestor up to / and refused the write if any was a
link, so macOS /var, a home on a linked volume, a dotfiles-linked ~/.pi, or a
PI_CODING_AGENT_DIR / CLAUDE_CONFIG_DIR / CODEX_HOME behind a link failed with
"refusing to modify symlinked native path".

noSymlink now takes the root the user chose (the Agent folder, the project,
or Skillshare's state dir) and checks only the path and the folders between
it and that root, refusing paths outside it. The root is compared lexically
because every checked path is built by joining onto it, so no EvalSymlinks
call is needed. A linked config file or a linked folder inside the root is
still refused, and Windows junctions are still detected via IsLinkMode.

* fix(plugin): bound the noSymlink ancestor walk by its relative depth

The walk stopped when p equalled root, but on Windows Dir spells a UNC
share root with a trailing separator (\\server\share\) while Clean keeps
the root without it, so the loop never ended for a PI_CODING_AGENT_DIR,
project root or StateDir on a share root. Walk up exactly as many levels
as the relative path has instead.
2026-10-04 02:50:00 +08:00
Willie 9788d486f7 feat(pi): edit extensions on any Pi at or above the verified minimum (#367)
* feat(pi): edit extensions on any Pi at or above the verified minimum

Pi 1.0.1 was released and every user who updated found the Extensions
tab read-only, because the gate matched an exact list of versions. The
package contract has not changed between patch releases, so the gate now
accepts any plain X.Y.Z at or above PiMinVersion (0.99.2), the oldest
version with executed evidence. Older, prerelease or unparsable versions
stay read-only.

The version matrix now also records Pi 1.0.1, which passes every check.
The evidence test requires PiMinVersion to be recorded and every recorded
version to pass, instead of matching an exact list.

The read-only reason is renamed to unsupportedVersion and the API returns
minVersion, so the dashboard says which version is needed.

The Pi MCP project switch keeps writing the global command or url: from
1.0.1 Pi accepts a switch-only entry, but it is skipped by older Pi and
warns at every start where the global config lacks the server. Its
rationale in comments and docs is updated to say so.

* fix(docs): update Pi extensions design records for the minimum gate

Codex review on #367: the design records still described the exact
version list (PiVerifiedVersions, unverifiedVersion, "exact versions
only"), so a contributor relying on them would implement or validate the
old contract. implementation-report.md now states the current gate and
evidence, and phase-0.md gains a dated minimum-version update that the
older paragraphs point to, keeping the history intact.
2026-10-04 02:26:59 +08:00
Willie 475b176914 fix(pull): keep the local config.yaml when a later pull brings a tracked copy (#362)
* fix(pull): keep the local config.yaml when a later pull brings a tracked copy

#357 refuses a root-scope first pull when the remote tracks config.yaml, but once upstream is set, a pull that merges in a remote-tracked copy still replaced this machine's config: Git treats the ignored file as expendable. Snapshot config.yaml around later pulls (CLI pull, push --pull, and the dashboard), restore it, and warn so the user can untrack it with push.

Refs #353

* fix(git): untrack config.yaml after push --pull and warn on identical copies

push --pull ran the root safety sweep before integrating the remote, so a
pull that brought in a remote-tracked config.yaml was pushed back as-is:
the remote kept tracking it and the next first pull elsewhere was refused.
Re-run EnsureConfigUntracked after the pull and commit the removal before
pushing.

The warning also depended on the pull changing the local bytes, so a
tracked copy identical to this machine's went unreported. KeepLocalConfig's
restore now reports whether the repo tracks config.yaml after the pull.

* fix(git): only restore config.yaml when the pull could have touched it

KeepLocalConfig wrote its snapshot back after every root-scope pull, so an
edit made to an untracked config.yaml while the pull ran was silently
replaced by the stale copy. Git only rewrites config.yaml when it is
tracked before or after the pull, so restore only in that case.

* fix(git): restore config.yaml only when the pull changed its index entry

Gating the restore on "tracked before or after" still overwrote edits made
during a pull when config.yaml was already tracked and the pull left it
alone. Compare config.yaml's index entry (mode, blob, stage) before and
after the pull instead; a merge, reset or conflict that touches the file
changes it, and anything else keeps the current file.

* fix(git): warn about a tracked config.yaml only when the upstream tracks it

The warning followed the local index, so a config.yaml tracked only by a
local, unpushed commit was reported as tracked by the remote, and the
suggested push would publish it in history. Check the upstream tree with
RemoteTracksConfig("@{u}") instead.

* fix(git): restore config.yaml after a pull that aborts its merge

A merge that brought in a tracked config.yaml and conflicted elsewhere was
aborted, which deleted the file again and returned the index entry to its
pre-pull state, so restore saw no change and the local config was lost.
Also restore when the snapshotted file is missing after the pull.

* fix(git): keep config.yaml ignored when a later rule negates it

EnsureConfigUntracked only checked that a "config.yaml" line existed, so a
pulled .gitignore with a later "!config.yaml" left the file unignored and
the following `git add -A` tracked this machine's copy again. Check the
effective ignore state and append an overriding entry when needed.

* fix: commit post-pull .gitignore repairs and keep config.yaml's mode

push --pull committed only when EnsureConfigUntracked removed an index
entry, so a pulled .gitignore that stopped ignoring config.yaml was
repaired locally but never pushed, leaving the tree dirty. Commit whenever
the safety step leaves changes.

KeepLocalConfig skipped the restore when the bytes matched, so a checkout
that rewrote config.yaml with the default mode dropped this machine's
restrictive permissions. Compare the mode too, and chmod after writing
since WriteFile's mode is masked by umask.

* fix(git): restore config.yaml over a tracked config.yaml directory

A pull that checks out a tracked config.yaml/ tree replaces the ignored
local file with a directory, which os.Remove cannot delete, so restore
failed and the local config stayed lost. Remove a directory there before
writing the snapshot back; its contents are in Git. EnsureConfigUntracked
now untracks recursively so push can remove that tree too.

* fix: keep post-checkout config edits and preview .gitignore repairs

Restore wrote the snapshot back whenever the pull changed config.yaml's
index entry, even when a newer copy had been written after Git checked it
out. Skip the restore when the file on disk no longer matches the blob the
pull checked out.

A dry-run push or commit only previewed untracking config.yaml, so when a
later "!config.yaml" rule un-ignored it the preview listed config.yaml to
commit while a real run repairs .gitignore instead. Preview that repair.

* fix: scope the post-pull safety commit and keep post-checkout symlinks

The post-pull safety step in push --pull ran `git add -A`, so an edit made
by another process while the pull ran was committed and pushed under the
config-safety message. Stage only .gitignore on top of the config.yaml
removal and commit the index as staged.

The post-checkout edit check skipped symlinks, so a dotfiles manager that
replaced config.yaml with a newer link after the checkout had it replaced
by the stale snapshot. Treat a link where Git checked out a file, or a
different link target, as a newer local edit.

* fix: preview explicit config.yaml entries and refuse other .gitignore edits

A dry run previewed the .gitignore repair only when config.yaml was not
ignored at all, but a real run also appends the explicit entry when a
broader rule such as *.yaml ignores it. Preview whenever the entry is
missing too.

The post-pull safety commit staged the whole .gitignore, so an edit made
by another process while the pull ran was pushed with it. Refuse unless
the file only gained config.yaml lines over its staged version.

* fix: report config.yaml restore failures instead of dropping them

integrateRemote declared the KeepLocalConfig result with :=, shadowing the
named err that the deferred restore joins its failure into, so a failed
restore was swallowed and pull or push --pull went on to sync targets with
another machine's config.yaml. The dashboard pull replaced the pull's own
error with a bare 500 on a restore failure, losing the conflict response
and the ops log entry; join the errors so both are reported.
2026-10-04 01:58:30 +08:00